Environmental regulation at European level compels member states to a better protection of wetland areas. Systematic inventory and delineations of wetlands have therefore been recently undertaken in France and consider soil redoximorphic features using criteria fixed by decree. Digital soil mapping is tested as a support of wetland delineation at the regional scale of Brittany (France) and the study focuses on two objectives: (i) to compare the efficiency of learning strategies based on punctual observations or existing detailed soil maps to extrapolate spatial models of soil waterlogging; (ii) to assess the influence of prediction uncertainties of soil waterlogging on the final wetland delineation. A classification tree is used: MART (Multiple Additive and Regression Trees). MART allows the creation of predictive model based on point data and predictive variables: topography, landscape map obtain by remote sensing, airborne gamma-ray spectrometry, etc. Two kinds of models are inferred: P-Model is based on 5129 punctual soil descriptions; M-Model is based on existing soil maps at 1:25 000 scale covering 11% of the study area. Model validation by independent data set indicates an overall better performance of P-Model explained by a better spatial coverage. M-Model nevertheless performed slightly better in areas similar to its learning conditions. Probability estimates of soil waterlogging classes were finally combined to estimate the probability of occurrence of wetland conditions meeting the regulation criteria.